Merge tag 'r8169-upstream-20061204-00' of git://electric-eye.fr.zoreil.com/home/romie...
[linux-2.6] / fs / ocfs2 / dlm / dlmrecovery.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * dlmrecovery.c
5  *
6  * recovery stuff
7  *
8  * Copyright (C) 2004 Oracle.  All rights reserved.
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public
12  * License as published by the Free Software Foundation; either
13  * version 2 of the License, or (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public
21  * License along with this program; if not, write to the
22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23  * Boston, MA 021110-1307, USA.
24  *
25  */
26
27
28 #include <linux/module.h>
29 #include <linux/fs.h>
30 #include <linux/types.h>
31 #include <linux/slab.h>
32 #include <linux/highmem.h>
33 #include <linux/utsname.h>
34 #include <linux/init.h>
35 #include <linux/sysctl.h>
36 #include <linux/random.h>
37 #include <linux/blkdev.h>
38 #include <linux/socket.h>
39 #include <linux/inet.h>
40 #include <linux/timer.h>
41 #include <linux/kthread.h>
42 #include <linux/delay.h>
43
44
45 #include "cluster/heartbeat.h"
46 #include "cluster/nodemanager.h"
47 #include "cluster/tcp.h"
48
49 #include "dlmapi.h"
50 #include "dlmcommon.h"
51 #include "dlmdomain.h"
52
53 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY)
54 #include "cluster/masklog.h"
55
56 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node);
57
58 static int dlm_recovery_thread(void *data);
59 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm);
60 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm);
61 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm);
62 static int dlm_do_recovery(struct dlm_ctxt *dlm);
63
64 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm);
65 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node);
66 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
67 static int dlm_request_all_locks(struct dlm_ctxt *dlm,
68                                  u8 request_from, u8 dead_node);
69 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
70
71 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res);
72 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
73                                         const char *lockname, int namelen,
74                                         int total_locks, u64 cookie,
75                                         u8 flags, u8 master);
76 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
77                                     struct dlm_migratable_lockres *mres,
78                                     u8 send_to,
79                                     struct dlm_lock_resource *res,
80                                     int total_locks);
81 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
82                                      struct dlm_lock_resource *res,
83                                      struct dlm_migratable_lockres *mres);
84 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm);
85 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm,
86                                  u8 dead_node, u8 send_to);
87 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node);
88 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
89                                         struct list_head *list, u8 dead_node);
90 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
91                                               u8 dead_node, u8 new_master);
92 static void dlm_reco_ast(void *astdata);
93 static void dlm_reco_bast(void *astdata, int blocked_type);
94 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st);
95 static void dlm_request_all_locks_worker(struct dlm_work_item *item,
96                                          void *data);
97 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data);
98 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
99                                       struct dlm_lock_resource *res,
100                                       u8 *real_master);
101
102 static u64 dlm_get_next_mig_cookie(void);
103
104 static DEFINE_SPINLOCK(dlm_reco_state_lock);
105 static DEFINE_SPINLOCK(dlm_mig_cookie_lock);
106 static u64 dlm_mig_cookie = 1;
107
108 static u64 dlm_get_next_mig_cookie(void)
109 {
110         u64 c;
111         spin_lock(&dlm_mig_cookie_lock);
112         c = dlm_mig_cookie;
113         if (dlm_mig_cookie == (~0ULL))
114                 dlm_mig_cookie = 1;
115         else
116                 dlm_mig_cookie++;
117         spin_unlock(&dlm_mig_cookie_lock);
118         return c;
119 }
120
121 static inline void dlm_set_reco_dead_node(struct dlm_ctxt *dlm,
122                                           u8 dead_node)
123 {
124         assert_spin_locked(&dlm->spinlock);
125         if (dlm->reco.dead_node != dead_node)
126                 mlog(0, "%s: changing dead_node from %u to %u\n",
127                      dlm->name, dlm->reco.dead_node, dead_node);
128         dlm->reco.dead_node = dead_node;
129 }
130
131 static inline void dlm_set_reco_master(struct dlm_ctxt *dlm,
132                                        u8 master)
133 {
134         assert_spin_locked(&dlm->spinlock);
135         mlog(0, "%s: changing new_master from %u to %u\n",
136              dlm->name, dlm->reco.new_master, master);
137         dlm->reco.new_master = master;
138 }
139
140 static inline void __dlm_reset_recovery(struct dlm_ctxt *dlm)
141 {
142         assert_spin_locked(&dlm->spinlock);
143         clear_bit(dlm->reco.dead_node, dlm->recovery_map);
144         dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
145         dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
146 }
147
148 static inline void dlm_reset_recovery(struct dlm_ctxt *dlm)
149 {
150         spin_lock(&dlm->spinlock);
151         __dlm_reset_recovery(dlm);
152         spin_unlock(&dlm->spinlock);
153 }
154
155 /* Worker function used during recovery. */
156 void dlm_dispatch_work(struct work_struct *work)
157 {
158         struct dlm_ctxt *dlm =
159                 container_of(work, struct dlm_ctxt, dispatched_work);
160         LIST_HEAD(tmp_list);
161         struct list_head *iter, *iter2;
162         struct dlm_work_item *item;
163         dlm_workfunc_t *workfunc;
164         int tot=0;
165
166         if (!dlm_joined(dlm))
167                 return;
168
169         spin_lock(&dlm->work_lock);
170         list_splice_init(&dlm->work_list, &tmp_list);
171         spin_unlock(&dlm->work_lock);
172
173         list_for_each_safe(iter, iter2, &tmp_list) {
174                 tot++;
175         }
176         mlog(0, "%s: work thread has %d work items\n", dlm->name, tot);
177
178         list_for_each_safe(iter, iter2, &tmp_list) {
179                 item = list_entry(iter, struct dlm_work_item, list);
180                 workfunc = item->func;
181                 list_del_init(&item->list);
182
183                 /* already have ref on dlm to avoid having
184                  * it disappear.  just double-check. */
185                 BUG_ON(item->dlm != dlm);
186
187                 /* this is allowed to sleep and
188                  * call network stuff */
189                 workfunc(item, item->data);
190
191                 dlm_put(dlm);
192                 kfree(item);
193         }
194 }
195
196 /*
197  * RECOVERY THREAD
198  */
199
200 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm)
201 {
202         /* wake the recovery thread
203          * this will wake the reco thread in one of three places
204          * 1) sleeping with no recovery happening
205          * 2) sleeping with recovery mastered elsewhere
206          * 3) recovery mastered here, waiting on reco data */
207
208         wake_up(&dlm->dlm_reco_thread_wq);
209 }
210
211 /* Launch the recovery thread */
212 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm)
213 {
214         mlog(0, "starting dlm recovery thread...\n");
215
216         dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm,
217                                                 "dlm_reco_thread");
218         if (IS_ERR(dlm->dlm_reco_thread_task)) {
219                 mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task));
220                 dlm->dlm_reco_thread_task = NULL;
221                 return -EINVAL;
222         }
223
224         return 0;
225 }
226
227 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm)
228 {
229         if (dlm->dlm_reco_thread_task) {
230                 mlog(0, "waiting for dlm recovery thread to exit\n");
231                 kthread_stop(dlm->dlm_reco_thread_task);
232                 dlm->dlm_reco_thread_task = NULL;
233         }
234 }
235
236
237
238 /*
239  * this is lame, but here's how recovery works...
240  * 1) all recovery threads cluster wide will work on recovering
241  *    ONE node at a time
242  * 2) negotiate who will take over all the locks for the dead node.
243  *    thats right... ALL the locks.
244  * 3) once a new master is chosen, everyone scans all locks
245  *    and moves aside those mastered by the dead guy
246  * 4) each of these locks should be locked until recovery is done
247  * 5) the new master collects up all of secondary lock queue info
248  *    one lock at a time, forcing each node to communicate back
249  *    before continuing
250  * 6) each secondary lock queue responds with the full known lock info
251  * 7) once the new master has run all its locks, it sends a ALLDONE!
252  *    message to everyone
253  * 8) upon receiving this message, the secondary queue node unlocks
254  *    and responds to the ALLDONE
255  * 9) once the new master gets responses from everyone, he unlocks
256  *    everything and recovery for this dead node is done
257  *10) go back to 2) while there are still dead nodes
258  *
259  */
260
261 static void dlm_print_reco_node_status(struct dlm_ctxt *dlm)
262 {
263         struct dlm_reco_node_data *ndata;
264         struct dlm_lock_resource *res;
265
266         mlog(ML_NOTICE, "%s(%d): recovery info, state=%s, dead=%u, master=%u\n",
267              dlm->name, dlm->dlm_reco_thread_task->pid,
268              dlm->reco.state & DLM_RECO_STATE_ACTIVE ? "ACTIVE" : "inactive",
269              dlm->reco.dead_node, dlm->reco.new_master);
270
271         list_for_each_entry(ndata, &dlm->reco.node_data, list) {
272                 char *st = "unknown";
273                 switch (ndata->state) {
274                         case DLM_RECO_NODE_DATA_INIT:
275                                 st = "init";
276                                 break;
277                         case DLM_RECO_NODE_DATA_REQUESTING:
278                                 st = "requesting";
279                                 break;
280                         case DLM_RECO_NODE_DATA_DEAD:
281                                 st = "dead";
282                                 break;
283                         case DLM_RECO_NODE_DATA_RECEIVING:
284                                 st = "receiving";
285                                 break;
286                         case DLM_RECO_NODE_DATA_REQUESTED:
287                                 st = "requested";
288                                 break;
289                         case DLM_RECO_NODE_DATA_DONE:
290                                 st = "done";
291                                 break;
292                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
293                                 st = "finalize-sent";
294                                 break;
295                         default:
296                                 st = "bad";
297                                 break;
298                 }
299                 mlog(ML_NOTICE, "%s: reco state, node %u, state=%s\n",
300                      dlm->name, ndata->node_num, st);
301         }
302         list_for_each_entry(res, &dlm->reco.resources, recovering) {
303                 mlog(ML_NOTICE, "%s: lockres %.*s on recovering list\n",
304                      dlm->name, res->lockname.len, res->lockname.name);
305         }
306 }
307
308 #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000)
309
310 static int dlm_recovery_thread(void *data)
311 {
312         int status;
313         struct dlm_ctxt *dlm = data;
314         unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS);
315
316         mlog(0, "dlm thread running for %s...\n", dlm->name);
317
318         while (!kthread_should_stop()) {
319                 if (dlm_joined(dlm)) {
320                         status = dlm_do_recovery(dlm);
321                         if (status == -EAGAIN) {
322                                 /* do not sleep, recheck immediately. */
323                                 continue;
324                         }
325                         if (status < 0)
326                                 mlog_errno(status);
327                 }
328
329                 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
330                                                  kthread_should_stop(),
331                                                  timeout);
332         }
333
334         mlog(0, "quitting DLM recovery thread\n");
335         return 0;
336 }
337
338 /* returns true when the recovery master has contacted us */
339 static int dlm_reco_master_ready(struct dlm_ctxt *dlm)
340 {
341         int ready;
342         spin_lock(&dlm->spinlock);
343         ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM);
344         spin_unlock(&dlm->spinlock);
345         return ready;
346 }
347
348 /* returns true if node is no longer in the domain
349  * could be dead or just not joined */
350 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node)
351 {
352         int dead;
353         spin_lock(&dlm->spinlock);
354         dead = !test_bit(node, dlm->domain_map);
355         spin_unlock(&dlm->spinlock);
356         return dead;
357 }
358
359 /* returns true if node is no longer in the domain
360  * could be dead or just not joined */
361 static int dlm_is_node_recovered(struct dlm_ctxt *dlm, u8 node)
362 {
363         int recovered;
364         spin_lock(&dlm->spinlock);
365         recovered = !test_bit(node, dlm->recovery_map);
366         spin_unlock(&dlm->spinlock);
367         return recovered;
368 }
369
370
371 int dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout)
372 {
373         if (timeout) {
374                 mlog(ML_NOTICE, "%s: waiting %dms for notification of "
375                      "death of node %u\n", dlm->name, timeout, node);
376                 wait_event_timeout(dlm->dlm_reco_thread_wq,
377                            dlm_is_node_dead(dlm, node),
378                            msecs_to_jiffies(timeout));
379         } else {
380                 mlog(ML_NOTICE, "%s: waiting indefinitely for notification "
381                      "of death of node %u\n", dlm->name, node);
382                 wait_event(dlm->dlm_reco_thread_wq,
383                            dlm_is_node_dead(dlm, node));
384         }
385         /* for now, return 0 */
386         return 0;
387 }
388
389 int dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout)
390 {
391         if (timeout) {
392                 mlog(0, "%s: waiting %dms for notification of "
393                      "recovery of node %u\n", dlm->name, timeout, node);
394                 wait_event_timeout(dlm->dlm_reco_thread_wq,
395                            dlm_is_node_recovered(dlm, node),
396                            msecs_to_jiffies(timeout));
397         } else {
398                 mlog(0, "%s: waiting indefinitely for notification "
399                      "of recovery of node %u\n", dlm->name, node);
400                 wait_event(dlm->dlm_reco_thread_wq,
401                            dlm_is_node_recovered(dlm, node));
402         }
403         /* for now, return 0 */
404         return 0;
405 }
406
407 /* callers of the top-level api calls (dlmlock/dlmunlock) should
408  * block on the dlm->reco.event when recovery is in progress.
409  * the dlm recovery thread will set this state when it begins
410  * recovering a dead node (as the new master or not) and clear
411  * the state and wake as soon as all affected lock resources have
412  * been marked with the RECOVERY flag */
413 static int dlm_in_recovery(struct dlm_ctxt *dlm)
414 {
415         int in_recovery;
416         spin_lock(&dlm->spinlock);
417         in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
418         spin_unlock(&dlm->spinlock);
419         return in_recovery;
420 }
421
422
423 void dlm_wait_for_recovery(struct dlm_ctxt *dlm)
424 {
425         if (dlm_in_recovery(dlm)) {
426                 mlog(0, "%s: reco thread %d in recovery: "
427                      "state=%d, master=%u, dead=%u\n",
428                      dlm->name, dlm->dlm_reco_thread_task->pid,
429                      dlm->reco.state, dlm->reco.new_master,
430                      dlm->reco.dead_node);
431         }
432         wait_event(dlm->reco.event, !dlm_in_recovery(dlm));
433 }
434
435 static void dlm_begin_recovery(struct dlm_ctxt *dlm)
436 {
437         spin_lock(&dlm->spinlock);
438         BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
439         dlm->reco.state |= DLM_RECO_STATE_ACTIVE;
440         spin_unlock(&dlm->spinlock);
441 }
442
443 static void dlm_end_recovery(struct dlm_ctxt *dlm)
444 {
445         spin_lock(&dlm->spinlock);
446         BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE));
447         dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE;
448         spin_unlock(&dlm->spinlock);
449         wake_up(&dlm->reco.event);
450 }
451
452 static int dlm_do_recovery(struct dlm_ctxt *dlm)
453 {
454         int status = 0;
455         int ret;
456
457         spin_lock(&dlm->spinlock);
458
459         /* check to see if the new master has died */
460         if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM &&
461             test_bit(dlm->reco.new_master, dlm->recovery_map)) {
462                 mlog(0, "new master %u died while recovering %u!\n",
463                      dlm->reco.new_master, dlm->reco.dead_node);
464                 /* unset the new_master, leave dead_node */
465                 dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
466         }
467
468         /* select a target to recover */
469         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
470                 int bit;
471
472                 bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES+1, 0);
473                 if (bit >= O2NM_MAX_NODES || bit < 0)
474                         dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
475                 else
476                         dlm_set_reco_dead_node(dlm, bit);
477         } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) {
478                 /* BUG? */
479                 mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n",
480                      dlm->reco.dead_node);
481                 dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
482         }
483
484         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
485                 // mlog(0, "nothing to recover!  sleeping now!\n");
486                 spin_unlock(&dlm->spinlock);
487                 /* return to main thread loop and sleep. */
488                 return 0;
489         }
490         mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n",
491              dlm->name, dlm->dlm_reco_thread_task->pid,
492              dlm->reco.dead_node);
493         spin_unlock(&dlm->spinlock);
494
495         /* take write barrier */
496         /* (stops the list reshuffling thread, proxy ast handling) */
497         dlm_begin_recovery(dlm);
498
499         if (dlm->reco.new_master == dlm->node_num)
500                 goto master_here;
501
502         if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
503                 /* choose a new master, returns 0 if this node
504                  * is the master, -EEXIST if it's another node.
505                  * this does not return until a new master is chosen
506                  * or recovery completes entirely. */
507                 ret = dlm_pick_recovery_master(dlm);
508                 if (!ret) {
509                         /* already notified everyone.  go. */
510                         goto master_here;
511                 }
512                 mlog(0, "another node will master this recovery session.\n");
513         }
514         mlog(0, "dlm=%s (%d), new_master=%u, this node=%u, dead_node=%u\n",
515              dlm->name, dlm->dlm_reco_thread_task->pid, dlm->reco.new_master,
516              dlm->node_num, dlm->reco.dead_node);
517
518         /* it is safe to start everything back up here
519          * because all of the dead node's lock resources
520          * have been marked as in-recovery */
521         dlm_end_recovery(dlm);
522
523         /* sleep out in main dlm_recovery_thread loop. */
524         return 0;
525
526 master_here:
527         mlog(0, "(%d) mastering recovery of %s:%u here(this=%u)!\n",
528              dlm->dlm_reco_thread_task->pid,
529              dlm->name, dlm->reco.dead_node, dlm->node_num);
530
531         status = dlm_remaster_locks(dlm, dlm->reco.dead_node);
532         if (status < 0) {
533                 /* we should never hit this anymore */
534                 mlog(ML_ERROR, "error %d remastering locks for node %u, "
535                      "retrying.\n", status, dlm->reco.dead_node);
536                 /* yield a bit to allow any final network messages
537                  * to get handled on remaining nodes */
538                 msleep(100);
539         } else {
540                 /* success!  see if any other nodes need recovery */
541                 mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n",
542                      dlm->name, dlm->reco.dead_node, dlm->node_num);
543                 dlm_reset_recovery(dlm);
544         }
545         dlm_end_recovery(dlm);
546
547         /* continue and look for another dead node */
548         return -EAGAIN;
549 }
550
551 static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
552 {
553         int status = 0;
554         struct dlm_reco_node_data *ndata;
555         struct list_head *iter;
556         int all_nodes_done;
557         int destroy = 0;
558         int pass = 0;
559
560         do {
561                 /* we have become recovery master.  there is no escaping
562                  * this, so just keep trying until we get it. */
563                 status = dlm_init_recovery_area(dlm, dead_node);
564                 if (status < 0) {
565                         mlog(ML_ERROR, "%s: failed to alloc recovery area, "
566                              "retrying\n", dlm->name);
567                         msleep(1000);
568                 }
569         } while (status != 0);
570
571         /* safe to access the node data list without a lock, since this
572          * process is the only one to change the list */
573         list_for_each(iter, &dlm->reco.node_data) {
574                 ndata = list_entry (iter, struct dlm_reco_node_data, list);
575                 BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT);
576                 ndata->state = DLM_RECO_NODE_DATA_REQUESTING;
577
578                 mlog(0, "requesting lock info from node %u\n",
579                      ndata->node_num);
580
581                 if (ndata->node_num == dlm->node_num) {
582                         ndata->state = DLM_RECO_NODE_DATA_DONE;
583                         continue;
584                 }
585
586                 do {
587                         status = dlm_request_all_locks(dlm, ndata->node_num,
588                                                        dead_node);
589                         if (status < 0) {
590                                 mlog_errno(status);
591                                 if (dlm_is_host_down(status)) {
592                                         /* node died, ignore it for recovery */
593                                         status = 0;
594                                         ndata->state = DLM_RECO_NODE_DATA_DEAD;
595                                         /* wait for the domain map to catch up
596                                          * with the network state. */
597                                         wait_event_timeout(dlm->dlm_reco_thread_wq,
598                                                            dlm_is_node_dead(dlm,
599                                                                 ndata->node_num),
600                                                            msecs_to_jiffies(1000));
601                                         mlog(0, "waited 1 sec for %u, "
602                                              "dead? %s\n", ndata->node_num,
603                                              dlm_is_node_dead(dlm, ndata->node_num) ?
604                                              "yes" : "no");
605                                 } else {
606                                         /* -ENOMEM on the other node */
607                                         mlog(0, "%s: node %u returned "
608                                              "%d during recovery, retrying "
609                                              "after a short wait\n",
610                                              dlm->name, ndata->node_num,
611                                              status);
612                                         msleep(100);
613                                 }
614                         }
615                 } while (status != 0);
616
617                 switch (ndata->state) {
618                         case DLM_RECO_NODE_DATA_INIT:
619                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
620                         case DLM_RECO_NODE_DATA_REQUESTED:
621                                 BUG();
622                                 break;
623                         case DLM_RECO_NODE_DATA_DEAD:
624                                 mlog(0, "node %u died after requesting "
625                                      "recovery info for node %u\n",
626                                      ndata->node_num, dead_node);
627                                 /* fine.  don't need this node's info.
628                                  * continue without it. */
629                                 break;
630                         case DLM_RECO_NODE_DATA_REQUESTING:
631                                 ndata->state = DLM_RECO_NODE_DATA_REQUESTED;
632                                 mlog(0, "now receiving recovery data from "
633                                      "node %u for dead node %u\n",
634                                      ndata->node_num, dead_node);
635                                 break;
636                         case DLM_RECO_NODE_DATA_RECEIVING:
637                                 mlog(0, "already receiving recovery data from "
638                                      "node %u for dead node %u\n",
639                                      ndata->node_num, dead_node);
640                                 break;
641                         case DLM_RECO_NODE_DATA_DONE:
642                                 mlog(0, "already DONE receiving recovery data "
643                                      "from node %u for dead node %u\n",
644                                      ndata->node_num, dead_node);
645                                 break;
646                 }
647         }
648
649         mlog(0, "done requesting all lock info\n");
650
651         /* nodes should be sending reco data now
652          * just need to wait */
653
654         while (1) {
655                 /* check all the nodes now to see if we are
656                  * done, or if anyone died */
657                 all_nodes_done = 1;
658                 spin_lock(&dlm_reco_state_lock);
659                 list_for_each(iter, &dlm->reco.node_data) {
660                         ndata = list_entry (iter, struct dlm_reco_node_data, list);
661
662                         mlog(0, "checking recovery state of node %u\n",
663                              ndata->node_num);
664                         switch (ndata->state) {
665                                 case DLM_RECO_NODE_DATA_INIT:
666                                 case DLM_RECO_NODE_DATA_REQUESTING:
667                                         mlog(ML_ERROR, "bad ndata state for "
668                                              "node %u: state=%d\n",
669                                              ndata->node_num, ndata->state);
670                                         BUG();
671                                         break;
672                                 case DLM_RECO_NODE_DATA_DEAD:
673                                         mlog(0, "node %u died after "
674                                              "requesting recovery info for "
675                                              "node %u\n", ndata->node_num,
676                                              dead_node);
677                                         break;
678                                 case DLM_RECO_NODE_DATA_RECEIVING:
679                                 case DLM_RECO_NODE_DATA_REQUESTED:
680                                         mlog(0, "%s: node %u still in state %s\n",
681                                              dlm->name, ndata->node_num,
682                                              ndata->state==DLM_RECO_NODE_DATA_RECEIVING ?
683                                              "receiving" : "requested");
684                                         all_nodes_done = 0;
685                                         break;
686                                 case DLM_RECO_NODE_DATA_DONE:
687                                         mlog(0, "%s: node %u state is done\n",
688                                              dlm->name, ndata->node_num);
689                                         break;
690                                 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
691                                         mlog(0, "%s: node %u state is finalize\n",
692                                              dlm->name, ndata->node_num);
693                                         break;
694                         }
695                 }
696                 spin_unlock(&dlm_reco_state_lock);
697
698                 mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass,
699                      all_nodes_done?"yes":"no");
700                 if (all_nodes_done) {
701                         int ret;
702
703                         /* all nodes are now in DLM_RECO_NODE_DATA_DONE state
704                          * just send a finalize message to everyone and
705                          * clean up */
706                         mlog(0, "all nodes are done! send finalize\n");
707                         ret = dlm_send_finalize_reco_message(dlm);
708                         if (ret < 0)
709                                 mlog_errno(ret);
710
711                         spin_lock(&dlm->spinlock);
712                         dlm_finish_local_lockres_recovery(dlm, dead_node,
713                                                           dlm->node_num);
714                         spin_unlock(&dlm->spinlock);
715                         mlog(0, "should be done with recovery!\n");
716
717                         mlog(0, "finishing recovery of %s at %lu, "
718                              "dead=%u, this=%u, new=%u\n", dlm->name,
719                              jiffies, dlm->reco.dead_node,
720                              dlm->node_num, dlm->reco.new_master);
721                         destroy = 1;
722                         status = 0;
723                         /* rescan everything marked dirty along the way */
724                         dlm_kick_thread(dlm, NULL);
725                         break;
726                 }
727                 /* wait to be signalled, with periodic timeout
728                  * to check for node death */
729                 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
730                                          kthread_should_stop(),
731                                          msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS));
732
733         }
734
735         if (destroy)
736                 dlm_destroy_recovery_area(dlm, dead_node);
737
738         mlog_exit(status);
739         return status;
740 }
741
742 static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
743 {
744         int num=0;
745         struct dlm_reco_node_data *ndata;
746
747         spin_lock(&dlm->spinlock);
748         memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map));
749         /* nodes can only be removed (by dying) after dropping
750          * this lock, and death will be trapped later, so this should do */
751         spin_unlock(&dlm->spinlock);
752
753         while (1) {
754                 num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num);
755                 if (num >= O2NM_MAX_NODES) {
756                         break;
757                 }
758                 BUG_ON(num == dead_node);
759
760                 ndata = kcalloc(1, sizeof(*ndata), GFP_NOFS);
761                 if (!ndata) {
762                         dlm_destroy_recovery_area(dlm, dead_node);
763                         return -ENOMEM;
764                 }
765                 ndata->node_num = num;
766                 ndata->state = DLM_RECO_NODE_DATA_INIT;
767                 spin_lock(&dlm_reco_state_lock);
768                 list_add_tail(&ndata->list, &dlm->reco.node_data);
769                 spin_unlock(&dlm_reco_state_lock);
770                 num++;
771         }
772
773         return 0;
774 }
775
776 static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
777 {
778         struct list_head *iter, *iter2;
779         struct dlm_reco_node_data *ndata;
780         LIST_HEAD(tmplist);
781
782         spin_lock(&dlm_reco_state_lock);
783         list_splice_init(&dlm->reco.node_data, &tmplist);
784         spin_unlock(&dlm_reco_state_lock);
785
786         list_for_each_safe(iter, iter2, &tmplist) {
787                 ndata = list_entry (iter, struct dlm_reco_node_data, list);
788                 list_del_init(&ndata->list);
789                 kfree(ndata);
790         }
791 }
792
793 static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from,
794                                  u8 dead_node)
795 {
796         struct dlm_lock_request lr;
797         enum dlm_status ret;
798
799         mlog(0, "\n");
800
801
802         mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
803                   "to %u\n", dead_node, request_from);
804
805         memset(&lr, 0, sizeof(lr));
806         lr.node_idx = dlm->node_num;
807         lr.dead_node = dead_node;
808
809         // send message
810         ret = DLM_NOLOCKMGR;
811         ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key,
812                                  &lr, sizeof(lr), request_from, NULL);
813
814         /* negative status is handled by caller */
815         if (ret < 0)
816                 mlog_errno(ret);
817
818         // return from here, then
819         // sleep until all received or error
820         return ret;
821
822 }
823
824 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data)
825 {
826         struct dlm_ctxt *dlm = data;
827         struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf;
828         char *buf = NULL;
829         struct dlm_work_item *item = NULL;
830
831         if (!dlm_grab(dlm))
832                 return -EINVAL;
833
834         if (lr->dead_node != dlm->reco.dead_node) {
835                 mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local "
836                      "dead_node is %u\n", dlm->name, lr->node_idx,
837                      lr->dead_node, dlm->reco.dead_node);
838                 dlm_print_reco_node_status(dlm);
839                 /* this is a hack */
840                 dlm_put(dlm);
841                 return -ENOMEM;
842         }
843         BUG_ON(lr->dead_node != dlm->reco.dead_node);
844
845         item = kcalloc(1, sizeof(*item), GFP_NOFS);
846         if (!item) {
847                 dlm_put(dlm);
848                 return -ENOMEM;
849         }
850
851         /* this will get freed by dlm_request_all_locks_worker */
852         buf = (char *) __get_free_page(GFP_NOFS);
853         if (!buf) {
854                 kfree(item);
855                 dlm_put(dlm);
856                 return -ENOMEM;
857         }
858
859         /* queue up work for dlm_request_all_locks_worker */
860         dlm_grab(dlm);  /* get an extra ref for the work item */
861         dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf);
862         item->u.ral.reco_master = lr->node_idx;
863         item->u.ral.dead_node = lr->dead_node;
864         spin_lock(&dlm->work_lock);
865         list_add_tail(&item->list, &dlm->work_list);
866         spin_unlock(&dlm->work_lock);
867         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
868
869         dlm_put(dlm);
870         return 0;
871 }
872
873 static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
874 {
875         struct dlm_migratable_lockres *mres;
876         struct dlm_lock_resource *res;
877         struct dlm_ctxt *dlm;
878         LIST_HEAD(resources);
879         struct list_head *iter;
880         int ret;
881         u8 dead_node, reco_master;
882         int skip_all_done = 0;
883
884         dlm = item->dlm;
885         dead_node = item->u.ral.dead_node;
886         reco_master = item->u.ral.reco_master;
887         mres = (struct dlm_migratable_lockres *)data;
888
889         mlog(0, "%s: recovery worker started, dead=%u, master=%u\n",
890              dlm->name, dead_node, reco_master);
891
892         if (dead_node != dlm->reco.dead_node ||
893             reco_master != dlm->reco.new_master) {
894                 /* worker could have been created before the recovery master
895                  * died.  if so, do not continue, but do not error. */
896                 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
897                         mlog(ML_NOTICE, "%s: will not send recovery state, "
898                              "recovery master %u died, thread=(dead=%u,mas=%u)"
899                              " current=(dead=%u,mas=%u)\n", dlm->name,
900                              reco_master, dead_node, reco_master,
901                              dlm->reco.dead_node, dlm->reco.new_master);
902                 } else {
903                         mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, "
904                              "master=%u), request(dead=%u, master=%u)\n",
905                              dlm->name, dlm->reco.dead_node,
906                              dlm->reco.new_master, dead_node, reco_master);
907                 }
908                 goto leave;
909         }
910
911         /* lock resources should have already been moved to the
912          * dlm->reco.resources list.  now move items from that list
913          * to a temp list if the dead owner matches.  note that the
914          * whole cluster recovers only one node at a time, so we
915          * can safely move UNKNOWN lock resources for each recovery
916          * session. */
917         dlm_move_reco_locks_to_list(dlm, &resources, dead_node);
918
919         /* now we can begin blasting lockreses without the dlm lock */
920
921         /* any errors returned will be due to the new_master dying,
922          * the dlm_reco_thread should detect this */
923         list_for_each(iter, &resources) {
924                 res = list_entry (iter, struct dlm_lock_resource, recovering);
925                 ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
926                                         DLM_MRES_RECOVERY);
927                 if (ret < 0) {
928                         mlog(ML_ERROR, "%s: node %u went down while sending "
929                              "recovery state for dead node %u, ret=%d\n", dlm->name,
930                              reco_master, dead_node, ret);
931                         skip_all_done = 1;
932                         break;
933                 }
934         }
935
936         /* move the resources back to the list */
937         spin_lock(&dlm->spinlock);
938         list_splice_init(&resources, &dlm->reco.resources);
939         spin_unlock(&dlm->spinlock);
940
941         if (!skip_all_done) {
942                 ret = dlm_send_all_done_msg(dlm, dead_node, reco_master);
943                 if (ret < 0) {
944                         mlog(ML_ERROR, "%s: node %u went down while sending "
945                              "recovery all-done for dead node %u, ret=%d\n",
946                              dlm->name, reco_master, dead_node, ret);
947                 }
948         }
949 leave:
950         free_page((unsigned long)data);
951 }
952
953
954 static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to)
955 {
956         int ret, tmpret;
957         struct dlm_reco_data_done done_msg;
958
959         memset(&done_msg, 0, sizeof(done_msg));
960         done_msg.node_idx = dlm->node_num;
961         done_msg.dead_node = dead_node;
962         mlog(0, "sending DATA DONE message to %u, "
963              "my node=%u, dead node=%u\n", send_to, done_msg.node_idx,
964              done_msg.dead_node);
965
966         ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg,
967                                  sizeof(done_msg), send_to, &tmpret);
968         if (ret < 0) {
969                 if (!dlm_is_host_down(ret)) {
970                         mlog_errno(ret);
971                         mlog(ML_ERROR, "%s: unknown error sending data-done "
972                              "to %u\n", dlm->name, send_to);
973                         BUG();
974                 }
975         } else
976                 ret = tmpret;
977         return ret;
978 }
979
980
981 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data)
982 {
983         struct dlm_ctxt *dlm = data;
984         struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf;
985         struct list_head *iter;
986         struct dlm_reco_node_data *ndata = NULL;
987         int ret = -EINVAL;
988
989         if (!dlm_grab(dlm))
990                 return -EINVAL;
991
992         mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
993              "node_idx=%u, this node=%u\n", done->dead_node,
994              dlm->reco.dead_node, done->node_idx, dlm->node_num);
995
996         mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node),
997                         "Got DATA DONE: dead_node=%u, reco.dead_node=%u, "
998                         "node_idx=%u, this node=%u\n", done->dead_node,
999                         dlm->reco.dead_node, done->node_idx, dlm->node_num);
1000
1001         spin_lock(&dlm_reco_state_lock);
1002         list_for_each(iter, &dlm->reco.node_data) {
1003                 ndata = list_entry (iter, struct dlm_reco_node_data, list);
1004                 if (ndata->node_num != done->node_idx)
1005                         continue;
1006
1007                 switch (ndata->state) {
1008                         /* should have moved beyond INIT but not to FINALIZE yet */
1009                         case DLM_RECO_NODE_DATA_INIT:
1010                         case DLM_RECO_NODE_DATA_DEAD:
1011                         case DLM_RECO_NODE_DATA_FINALIZE_SENT:
1012                                 mlog(ML_ERROR, "bad ndata state for node %u:"
1013                                      " state=%d\n", ndata->node_num,
1014                                      ndata->state);
1015                                 BUG();
1016                                 break;
1017                         /* these states are possible at this point, anywhere along
1018                          * the line of recovery */
1019                         case DLM_RECO_NODE_DATA_DONE:
1020                         case DLM_RECO_NODE_DATA_RECEIVING:
1021                         case DLM_RECO_NODE_DATA_REQUESTED:
1022                         case DLM_RECO_NODE_DATA_REQUESTING:
1023                                 mlog(0, "node %u is DONE sending "
1024                                           "recovery data!\n",
1025                                           ndata->node_num);
1026
1027                                 ndata->state = DLM_RECO_NODE_DATA_DONE;
1028                                 ret = 0;
1029                                 break;
1030                 }
1031         }
1032         spin_unlock(&dlm_reco_state_lock);
1033
1034         /* wake the recovery thread, some node is done */
1035         if (!ret)
1036                 dlm_kick_recovery_thread(dlm);
1037
1038         if (ret < 0)
1039                 mlog(ML_ERROR, "failed to find recovery node data for node "
1040                      "%u\n", done->node_idx);
1041         dlm_put(dlm);
1042
1043         mlog(0, "leaving reco data done handler, ret=%d\n", ret);
1044         return ret;
1045 }
1046
1047 static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
1048                                         struct list_head *list,
1049                                         u8 dead_node)
1050 {
1051         struct dlm_lock_resource *res;
1052         struct list_head *iter, *iter2;
1053         struct dlm_lock *lock;
1054
1055         spin_lock(&dlm->spinlock);
1056         list_for_each_safe(iter, iter2, &dlm->reco.resources) {
1057                 res = list_entry (iter, struct dlm_lock_resource, recovering);
1058                 /* always prune any $RECOVERY entries for dead nodes,
1059                  * otherwise hangs can occur during later recovery */
1060                 if (dlm_is_recovery_lock(res->lockname.name,
1061                                          res->lockname.len)) {
1062                         spin_lock(&res->spinlock);
1063                         list_for_each_entry(lock, &res->granted, list) {
1064                                 if (lock->ml.node == dead_node) {
1065                                         mlog(0, "AHA! there was "
1066                                              "a $RECOVERY lock for dead "
1067                                              "node %u (%s)!\n", 
1068                                              dead_node, dlm->name);
1069                                         list_del_init(&lock->list);
1070                                         dlm_lock_put(lock);
1071                                         break;
1072                                 }
1073                         }
1074                         spin_unlock(&res->spinlock);
1075                         continue;
1076                 }
1077
1078                 if (res->owner == dead_node) {
1079                         mlog(0, "found lockres owned by dead node while "
1080                                   "doing recovery for node %u. sending it.\n",
1081                                   dead_node);
1082                         list_move_tail(&res->recovering, list);
1083                 } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
1084                         mlog(0, "found UNKNOWN owner while doing recovery "
1085                                   "for node %u. sending it.\n", dead_node);
1086                         list_move_tail(&res->recovering, list);
1087                 }
1088         }
1089         spin_unlock(&dlm->spinlock);
1090 }
1091
1092 static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res)
1093 {
1094         int total_locks = 0;
1095         struct list_head *iter, *queue = &res->granted;
1096         int i;
1097
1098         for (i=0; i<3; i++) {
1099                 list_for_each(iter, queue)
1100                         total_locks++;
1101                 queue++;
1102         }
1103         return total_locks;
1104 }
1105
1106
1107 static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
1108                                       struct dlm_migratable_lockres *mres,
1109                                       u8 send_to,
1110                                       struct dlm_lock_resource *res,
1111                                       int total_locks)
1112 {
1113         u64 mig_cookie = be64_to_cpu(mres->mig_cookie);
1114         int mres_total_locks = be32_to_cpu(mres->total_locks);
1115         int sz, ret = 0, status = 0;
1116         u8 orig_flags = mres->flags,
1117            orig_master = mres->master;
1118
1119         BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS);
1120         if (!mres->num_locks)
1121                 return 0;
1122
1123         sz = sizeof(struct dlm_migratable_lockres) +
1124                 (mres->num_locks * sizeof(struct dlm_migratable_lock));
1125
1126         /* add an all-done flag if we reached the last lock */
1127         orig_flags = mres->flags;
1128         BUG_ON(total_locks > mres_total_locks);
1129         if (total_locks == mres_total_locks)
1130                 mres->flags |= DLM_MRES_ALL_DONE;
1131
1132         /* send it */
1133         ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres,
1134                                  sz, send_to, &status);
1135         if (ret < 0) {
1136                 /* XXX: negative status is not handled.
1137                  * this will end up killing this node. */
1138                 mlog_errno(ret);
1139         } else {
1140                 /* might get an -ENOMEM back here */
1141                 ret = status;
1142                 if (ret < 0) {
1143                         mlog_errno(ret);
1144
1145                         if (ret == -EFAULT) {
1146                                 mlog(ML_ERROR, "node %u told me to kill "
1147                                      "myself!\n", send_to);
1148                                 BUG();
1149                         }
1150                 }
1151         }
1152
1153         /* zero and reinit the message buffer */
1154         dlm_init_migratable_lockres(mres, res->lockname.name,
1155                                     res->lockname.len, mres_total_locks,
1156                                     mig_cookie, orig_flags, orig_master);
1157         return ret;
1158 }
1159
1160 static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
1161                                         const char *lockname, int namelen,
1162                                         int total_locks, u64 cookie,
1163                                         u8 flags, u8 master)
1164 {
1165         /* mres here is one full page */
1166         memset(mres, 0, PAGE_SIZE);
1167         mres->lockname_len = namelen;
1168         memcpy(mres->lockname, lockname, namelen);
1169         mres->num_locks = 0;
1170         mres->total_locks = cpu_to_be32(total_locks);
1171         mres->mig_cookie = cpu_to_be64(cookie);
1172         mres->flags = flags;
1173         mres->master = master;
1174 }
1175
1176
1177 /* returns 1 if this lock fills the network structure,
1178  * 0 otherwise */
1179 static int dlm_add_lock_to_array(struct dlm_lock *lock,
1180                                  struct dlm_migratable_lockres *mres, int queue)
1181 {
1182         struct dlm_migratable_lock *ml;
1183         int lock_num = mres->num_locks;
1184
1185         ml = &(mres->ml[lock_num]);
1186         ml->cookie = lock->ml.cookie;
1187         ml->type = lock->ml.type;
1188         ml->convert_type = lock->ml.convert_type;
1189         ml->highest_blocked = lock->ml.highest_blocked;
1190         ml->list = queue;
1191         if (lock->lksb) {
1192                 ml->flags = lock->lksb->flags;
1193                 /* send our current lvb */
1194                 if (ml->type == LKM_EXMODE ||
1195                     ml->type == LKM_PRMODE) {
1196                         /* if it is already set, this had better be a PR
1197                          * and it has to match */
1198                         if (!dlm_lvb_is_empty(mres->lvb) &&
1199                             (ml->type == LKM_EXMODE ||
1200                              memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))) {
1201                                 mlog(ML_ERROR, "mismatched lvbs!\n");
1202                                 __dlm_print_one_lock_resource(lock->lockres);
1203                                 BUG();
1204                         }
1205                         memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN);
1206                 }
1207         }
1208         ml->node = lock->ml.node;
1209         mres->num_locks++;
1210         /* we reached the max, send this network message */
1211         if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS)
1212                 return 1;
1213         return 0;
1214 }
1215
1216
1217 int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1218                          struct dlm_migratable_lockres *mres,
1219                          u8 send_to, u8 flags)
1220 {
1221         struct list_head *queue, *iter;
1222         int total_locks, i;
1223         u64 mig_cookie = 0;
1224         struct dlm_lock *lock;
1225         int ret = 0;
1226
1227         BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1228
1229         mlog(0, "sending to %u\n", send_to);
1230
1231         total_locks = dlm_num_locks_in_lockres(res);
1232         if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) {
1233                 /* rare, but possible */
1234                 mlog(0, "argh.  lockres has %d locks.  this will "
1235                           "require more than one network packet to "
1236                           "migrate\n", total_locks);
1237                 mig_cookie = dlm_get_next_mig_cookie();
1238         }
1239
1240         dlm_init_migratable_lockres(mres, res->lockname.name,
1241                                     res->lockname.len, total_locks,
1242                                     mig_cookie, flags, res->owner);
1243
1244         total_locks = 0;
1245         for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) {
1246                 queue = dlm_list_idx_to_ptr(res, i);
1247                 list_for_each(iter, queue) {
1248                         lock = list_entry (iter, struct dlm_lock, list);
1249
1250                         /* add another lock. */
1251                         total_locks++;
1252                         if (!dlm_add_lock_to_array(lock, mres, i))
1253                                 continue;
1254
1255                         /* this filled the lock message,
1256                          * we must send it immediately. */
1257                         ret = dlm_send_mig_lockres_msg(dlm, mres, send_to,
1258                                                        res, total_locks);
1259                         if (ret < 0)
1260                                 goto error;
1261                 }
1262         }
1263         /* flush any remaining locks */
1264         ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks);
1265         if (ret < 0)
1266                 goto error;
1267         return ret;
1268
1269 error:
1270         mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n",
1271              dlm->name, ret);
1272         if (!dlm_is_host_down(ret))
1273                 BUG();
1274         mlog(0, "%s: node %u went down while sending %s "
1275              "lockres %.*s\n", dlm->name, send_to,
1276              flags & DLM_MRES_RECOVERY ?  "recovery" : "migration",
1277              res->lockname.len, res->lockname.name);
1278         return ret;
1279 }
1280
1281
1282
1283 /*
1284  * this message will contain no more than one page worth of
1285  * recovery data, and it will work on only one lockres.
1286  * there may be many locks in this page, and we may need to wait
1287  * for additional packets to complete all the locks (rare, but
1288  * possible).
1289  */
1290 /*
1291  * NOTE: the allocation error cases here are scary
1292  * we really cannot afford to fail an alloc in recovery
1293  * do we spin?  returning an error only delays the problem really
1294  */
1295
1296 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data)
1297 {
1298         struct dlm_ctxt *dlm = data;
1299         struct dlm_migratable_lockres *mres =
1300                 (struct dlm_migratable_lockres *)msg->buf;
1301         int ret = 0;
1302         u8 real_master;
1303         char *buf = NULL;
1304         struct dlm_work_item *item = NULL;
1305         struct dlm_lock_resource *res = NULL;
1306
1307         if (!dlm_grab(dlm))
1308                 return -EINVAL;
1309
1310         BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1311
1312         real_master = mres->master;
1313         if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1314                 /* cannot migrate a lockres with no master */
1315                 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1316         }
1317
1318         mlog(0, "%s message received from node %u\n",
1319                   (mres->flags & DLM_MRES_RECOVERY) ?
1320                   "recovery" : "migration", mres->master);
1321         if (mres->flags & DLM_MRES_ALL_DONE)
1322                 mlog(0, "all done flag.  all lockres data received!\n");
1323
1324         ret = -ENOMEM;
1325         buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS);
1326         item = kcalloc(1, sizeof(*item), GFP_NOFS);
1327         if (!buf || !item)
1328                 goto leave;
1329
1330         /* lookup the lock to see if we have a secondary queue for this
1331          * already...  just add the locks in and this will have its owner
1332          * and RECOVERY flag changed when it completes. */
1333         res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len);
1334         if (res) {
1335                 /* this will get a ref on res */
1336                 /* mark it as recovering/migrating and hash it */
1337                 spin_lock(&res->spinlock);
1338                 if (mres->flags & DLM_MRES_RECOVERY) {
1339                         res->state |= DLM_LOCK_RES_RECOVERING;
1340                 } else {
1341                         if (res->state & DLM_LOCK_RES_MIGRATING) {
1342                                 /* this is at least the second
1343                                  * lockres message */
1344                                 mlog(0, "lock %.*s is already migrating\n",
1345                                           mres->lockname_len,
1346                                           mres->lockname);
1347                         } else if (res->state & DLM_LOCK_RES_RECOVERING) {
1348                                 /* caller should BUG */
1349                                 mlog(ML_ERROR, "node is attempting to migrate "
1350                                      "lock %.*s, but marked as recovering!\n",
1351                                      mres->lockname_len, mres->lockname);
1352                                 ret = -EFAULT;
1353                                 spin_unlock(&res->spinlock);
1354                                 goto leave;
1355                         }
1356                         res->state |= DLM_LOCK_RES_MIGRATING;
1357                 }
1358                 spin_unlock(&res->spinlock);
1359         } else {
1360                 /* need to allocate, just like if it was
1361                  * mastered here normally  */
1362                 res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len);
1363                 if (!res)
1364                         goto leave;
1365
1366                 /* to match the ref that we would have gotten if
1367                  * dlm_lookup_lockres had succeeded */
1368                 dlm_lockres_get(res);
1369
1370                 /* mark it as recovering/migrating and hash it */
1371                 if (mres->flags & DLM_MRES_RECOVERY)
1372                         res->state |= DLM_LOCK_RES_RECOVERING;
1373                 else
1374                         res->state |= DLM_LOCK_RES_MIGRATING;
1375
1376                 spin_lock(&dlm->spinlock);
1377                 __dlm_insert_lockres(dlm, res);
1378                 spin_unlock(&dlm->spinlock);
1379
1380                 /* now that the new lockres is inserted,
1381                  * make it usable by other processes */
1382                 spin_lock(&res->spinlock);
1383                 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1384                 spin_unlock(&res->spinlock);
1385
1386                 /* add an extra ref for just-allocated lockres 
1387                  * otherwise the lockres will be purged immediately */
1388                 dlm_lockres_get(res);
1389
1390         }
1391
1392         /* at this point we have allocated everything we need,
1393          * and we have a hashed lockres with an extra ref and
1394          * the proper res->state flags. */
1395         ret = 0;
1396         if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1397                 /* migration cannot have an unknown master */
1398                 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1399                 mlog(0, "recovery has passed me a lockres with an "
1400                           "unknown owner.. will need to requery: "
1401                           "%.*s\n", mres->lockname_len, mres->lockname);
1402         } else {
1403                 spin_lock(&res->spinlock);
1404                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
1405                 spin_unlock(&res->spinlock);
1406         }
1407
1408         /* queue up work for dlm_mig_lockres_worker */
1409         dlm_grab(dlm);  /* get an extra ref for the work item */
1410         memcpy(buf, msg->buf, be16_to_cpu(msg->data_len));  /* copy the whole message */
1411         dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf);
1412         item->u.ml.lockres = res; /* already have a ref */
1413         item->u.ml.real_master = real_master;
1414         spin_lock(&dlm->work_lock);
1415         list_add_tail(&item->list, &dlm->work_list);
1416         spin_unlock(&dlm->work_lock);
1417         queue_work(dlm->dlm_worker, &dlm->dispatched_work);
1418
1419 leave:
1420         dlm_put(dlm);
1421         if (ret < 0) {
1422                 if (buf)
1423                         kfree(buf);
1424                 if (item)
1425                         kfree(item);
1426         }
1427
1428         mlog_exit(ret);
1429         return ret;
1430 }
1431
1432
1433 static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data)
1434 {
1435         struct dlm_ctxt *dlm = data;
1436         struct dlm_migratable_lockres *mres;
1437         int ret = 0;
1438         struct dlm_lock_resource *res;
1439         u8 real_master;
1440
1441         dlm = item->dlm;
1442         mres = (struct dlm_migratable_lockres *)data;
1443
1444         res = item->u.ml.lockres;
1445         real_master = item->u.ml.real_master;
1446
1447         if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1448                 /* this case is super-rare. only occurs if
1449                  * node death happens during migration. */
1450 again:
1451                 ret = dlm_lockres_master_requery(dlm, res, &real_master);
1452                 if (ret < 0) {
1453                         mlog(0, "dlm_lockres_master_requery ret=%d\n",
1454                                   ret);
1455                         goto again;
1456                 }
1457                 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1458                         mlog(0, "lockres %.*s not claimed.  "
1459                                    "this node will take it.\n",
1460                                    res->lockname.len, res->lockname.name);
1461                 } else {
1462                         mlog(0, "master needs to respond to sender "
1463                                   "that node %u still owns %.*s\n",
1464                                   real_master, res->lockname.len,
1465                                   res->lockname.name);
1466                         /* cannot touch this lockres */
1467                         goto leave;
1468                 }
1469         }
1470
1471         ret = dlm_process_recovery_data(dlm, res, mres);
1472         if (ret < 0)
1473                 mlog(0, "dlm_process_recovery_data returned  %d\n", ret);
1474         else
1475                 mlog(0, "dlm_process_recovery_data succeeded\n");
1476
1477         if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) ==
1478                            (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) {
1479                 ret = dlm_finish_migration(dlm, res, mres->master);
1480                 if (ret < 0)
1481                         mlog_errno(ret);
1482         }
1483
1484 leave:
1485         kfree(data);
1486         mlog_exit(ret);
1487 }
1488
1489
1490
1491 static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
1492                                       struct dlm_lock_resource *res,
1493                                       u8 *real_master)
1494 {
1495         struct dlm_node_iter iter;
1496         int nodenum;
1497         int ret = 0;
1498
1499         *real_master = DLM_LOCK_RES_OWNER_UNKNOWN;
1500
1501         /* we only reach here if one of the two nodes in a
1502          * migration died while the migration was in progress.
1503          * at this point we need to requery the master.  we
1504          * know that the new_master got as far as creating
1505          * an mle on at least one node, but we do not know
1506          * if any nodes had actually cleared the mle and set
1507          * the master to the new_master.  the old master
1508          * is supposed to set the owner to UNKNOWN in the
1509          * event of a new_master death, so the only possible
1510          * responses that we can get from nodes here are
1511          * that the master is new_master, or that the master
1512          * is UNKNOWN.
1513          * if all nodes come back with UNKNOWN then we know
1514          * the lock needs remastering here.
1515          * if any node comes back with a valid master, check
1516          * to see if that master is the one that we are
1517          * recovering.  if so, then the new_master died and
1518          * we need to remaster this lock.  if not, then the
1519          * new_master survived and that node will respond to
1520          * other nodes about the owner.
1521          * if there is an owner, this node needs to dump this
1522          * lockres and alert the sender that this lockres
1523          * was rejected. */
1524         spin_lock(&dlm->spinlock);
1525         dlm_node_iter_init(dlm->domain_map, &iter);
1526         spin_unlock(&dlm->spinlock);
1527
1528         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1529                 /* do not send to self */
1530                 if (nodenum == dlm->node_num)
1531                         continue;
1532                 ret = dlm_do_master_requery(dlm, res, nodenum, real_master);
1533                 if (ret < 0) {
1534                         mlog_errno(ret);
1535                         if (!dlm_is_host_down(ret))
1536                                 BUG();
1537                         /* host is down, so answer for that node would be
1538                          * DLM_LOCK_RES_OWNER_UNKNOWN.  continue. */
1539                 }
1540                 if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1541                         mlog(0, "lock master is %u\n", *real_master);
1542                         break;
1543                 }
1544         }
1545         return ret;
1546 }
1547
1548
1549 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1550                           u8 nodenum, u8 *real_master)
1551 {
1552         int ret = -EINVAL;
1553         struct dlm_master_requery req;
1554         int status = DLM_LOCK_RES_OWNER_UNKNOWN;
1555
1556         memset(&req, 0, sizeof(req));
1557         req.node_idx = dlm->node_num;
1558         req.namelen = res->lockname.len;
1559         memcpy(req.name, res->lockname.name, res->lockname.len);
1560
1561         ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key,
1562                                  &req, sizeof(req), nodenum, &status);
1563         /* XXX: negative status not handled properly here. */
1564         if (ret < 0)
1565                 mlog_errno(ret);
1566         else {
1567                 BUG_ON(status < 0);
1568                 BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN);
1569                 *real_master = (u8) (status & 0xff);
1570                 mlog(0, "node %u responded to master requery with %u\n",
1571                           nodenum, *real_master);
1572                 ret = 0;
1573         }
1574         return ret;
1575 }
1576
1577
1578 /* this function cannot error, so unless the sending
1579  * or receiving of the message failed, the owner can
1580  * be trusted */
1581 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data)
1582 {
1583         struct dlm_ctxt *dlm = data;
1584         struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf;
1585         struct dlm_lock_resource *res = NULL;
1586         unsigned int hash;
1587         int master = DLM_LOCK_RES_OWNER_UNKNOWN;
1588         u32 flags = DLM_ASSERT_MASTER_REQUERY;
1589
1590         if (!dlm_grab(dlm)) {
1591                 /* since the domain has gone away on this
1592                  * node, the proper response is UNKNOWN */
1593                 return master;
1594         }
1595
1596         hash = dlm_lockid_hash(req->name, req->namelen);
1597
1598         spin_lock(&dlm->spinlock);
1599         res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash);
1600         if (res) {
1601                 spin_lock(&res->spinlock);
1602                 master = res->owner;
1603                 if (master == dlm->node_num) {
1604                         int ret = dlm_dispatch_assert_master(dlm, res,
1605                                                              0, 0, flags);
1606                         if (ret < 0) {
1607                                 mlog_errno(-ENOMEM);
1608                                 /* retry!? */
1609                                 BUG();
1610                         }
1611                 }
1612                 spin_unlock(&res->spinlock);
1613         }
1614         spin_unlock(&dlm->spinlock);
1615
1616         dlm_put(dlm);
1617         return master;
1618 }
1619
1620 static inline struct list_head *
1621 dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num)
1622 {
1623         struct list_head *ret;
1624         BUG_ON(list_num < 0);
1625         BUG_ON(list_num > 2);
1626         ret = &(res->granted);
1627         ret += list_num;
1628         return ret;
1629 }
1630 /* TODO: do ast flush business
1631  * TODO: do MIGRATING and RECOVERING spinning
1632  */
1633
1634 /*
1635 * NOTE about in-flight requests during migration:
1636 *
1637 * Before attempting the migrate, the master has marked the lockres as
1638 * MIGRATING and then flushed all of its pending ASTS.  So any in-flight
1639 * requests either got queued before the MIGRATING flag got set, in which
1640 * case the lock data will reflect the change and a return message is on
1641 * the way, or the request failed to get in before MIGRATING got set.  In
1642 * this case, the caller will be told to spin and wait for the MIGRATING
1643 * flag to be dropped, then recheck the master.
1644 * This holds true for the convert, cancel and unlock cases, and since lvb
1645 * updates are tied to these same messages, it applies to lvb updates as
1646 * well.  For the lock case, there is no way a lock can be on the master
1647 * queue and not be on the secondary queue since the lock is always added
1648 * locally first.  This means that the new target node will never be sent
1649 * a lock that he doesn't already have on the list.
1650 * In total, this means that the local lock is correct and should not be
1651 * updated to match the one sent by the master.  Any messages sent back
1652 * from the master before the MIGRATING flag will bring the lock properly
1653 * up-to-date, and the change will be ordered properly for the waiter.
1654 * We will *not* attempt to modify the lock underneath the waiter.
1655 */
1656
1657 static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
1658                                      struct dlm_lock_resource *res,
1659                                      struct dlm_migratable_lockres *mres)
1660 {
1661         struct dlm_migratable_lock *ml;
1662         struct list_head *queue;
1663         struct dlm_lock *newlock = NULL;
1664         struct dlm_lockstatus *lksb = NULL;
1665         int ret = 0;
1666         int i, bad;
1667         struct list_head *iter;
1668         struct dlm_lock *lock = NULL;
1669
1670         mlog(0, "running %d locks for this lockres\n", mres->num_locks);
1671         for (i=0; i<mres->num_locks; i++) {
1672                 ml = &(mres->ml[i]);
1673                 BUG_ON(ml->highest_blocked != LKM_IVMODE);
1674                 newlock = NULL;
1675                 lksb = NULL;
1676
1677                 queue = dlm_list_num_to_pointer(res, ml->list);
1678
1679                 /* if the lock is for the local node it needs to
1680                  * be moved to the proper location within the queue.
1681                  * do not allocate a new lock structure. */
1682                 if (ml->node == dlm->node_num) {
1683                         /* MIGRATION ONLY! */
1684                         BUG_ON(!(mres->flags & DLM_MRES_MIGRATION));
1685
1686                         spin_lock(&res->spinlock);
1687                         list_for_each(iter, queue) {
1688                                 lock = list_entry (iter, struct dlm_lock, list);
1689                                 if (lock->ml.cookie != ml->cookie)
1690                                         lock = NULL;
1691                                 else
1692                                         break;
1693                         }
1694
1695                         /* lock is always created locally first, and
1696                          * destroyed locally last.  it must be on the list */
1697                         if (!lock) {
1698                                 u64 c = ml->cookie;
1699                                 mlog(ML_ERROR, "could not find local lock "
1700                                                "with cookie %u:%llu!\n",
1701                                                dlm_get_lock_cookie_node(c),
1702                                                dlm_get_lock_cookie_seq(c));
1703                                 BUG();
1704                         }
1705                         BUG_ON(lock->ml.node != ml->node);
1706
1707                         /* see NOTE above about why we do not update
1708                          * to match the master here */
1709
1710                         /* move the lock to its proper place */
1711                         /* do not alter lock refcount.  switching lists. */
1712                         list_move_tail(&lock->list, queue);
1713                         spin_unlock(&res->spinlock);
1714
1715                         mlog(0, "just reordered a local lock!\n");
1716                         continue;
1717                 }
1718
1719                 /* lock is for another node. */
1720                 newlock = dlm_new_lock(ml->type, ml->node,
1721                                        be64_to_cpu(ml->cookie), NULL);
1722                 if (!newlock) {
1723                         ret = -ENOMEM;
1724                         goto leave;
1725                 }
1726                 lksb = newlock->lksb;
1727                 dlm_lock_attach_lockres(newlock, res);
1728
1729                 if (ml->convert_type != LKM_IVMODE) {
1730                         BUG_ON(queue != &res->converting);
1731                         newlock->ml.convert_type = ml->convert_type;
1732                 }
1733                 lksb->flags |= (ml->flags &
1734                                 (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
1735
1736                 if (ml->type == LKM_NLMODE)
1737                         goto skip_lvb;
1738
1739                 if (!dlm_lvb_is_empty(mres->lvb)) {
1740                         if (lksb->flags & DLM_LKSB_PUT_LVB) {
1741                                 /* other node was trying to update
1742                                  * lvb when node died.  recreate the
1743                                  * lksb with the updated lvb. */
1744                                 memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN);
1745                                 /* the lock resource lvb update must happen
1746                                  * NOW, before the spinlock is dropped.
1747                                  * we no longer wait for the AST to update
1748                                  * the lvb. */
1749                                 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1750                         } else {
1751                                 /* otherwise, the node is sending its 
1752                                  * most recent valid lvb info */
1753                                 BUG_ON(ml->type != LKM_EXMODE &&
1754                                        ml->type != LKM_PRMODE);
1755                                 if (!dlm_lvb_is_empty(res->lvb) &&
1756                                     (ml->type == LKM_EXMODE ||
1757                                      memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) {
1758                                         int i;
1759                                         mlog(ML_ERROR, "%s:%.*s: received bad "
1760                                              "lvb! type=%d\n", dlm->name,
1761                                              res->lockname.len,
1762                                              res->lockname.name, ml->type);
1763                                         printk("lockres lvb=[");
1764                                         for (i=0; i<DLM_LVB_LEN; i++)
1765                                                 printk("%02x", res->lvb[i]);
1766                                         printk("]\nmigrated lvb=[");
1767                                         for (i=0; i<DLM_LVB_LEN; i++)
1768                                                 printk("%02x", mres->lvb[i]);
1769                                         printk("]\n");
1770                                         dlm_print_one_lock_resource(res);
1771                                         BUG();
1772                                 }
1773                                 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1774                         }
1775                 }
1776 skip_lvb:
1777
1778                 /* NOTE:
1779                  * wrt lock queue ordering and recovery:
1780                  *    1. order of locks on granted queue is
1781                  *       meaningless.
1782                  *    2. order of locks on converting queue is
1783                  *       LOST with the node death.  sorry charlie.
1784                  *    3. order of locks on the blocked queue is
1785                  *       also LOST.
1786                  * order of locks does not affect integrity, it
1787                  * just means that a lock request may get pushed
1788                  * back in line as a result of the node death.
1789                  * also note that for a given node the lock order
1790                  * for its secondary queue locks is preserved
1791                  * relative to each other, but clearly *not*
1792                  * preserved relative to locks from other nodes.
1793                  */
1794                 bad = 0;
1795                 spin_lock(&res->spinlock);
1796                 list_for_each_entry(lock, queue, list) {
1797                         if (lock->ml.cookie == ml->cookie) {
1798                                 u64 c = lock->ml.cookie;
1799                                 mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already "
1800                                      "exists on this lockres!\n", dlm->name,
1801                                      res->lockname.len, res->lockname.name,
1802                                      dlm_get_lock_cookie_node(c),
1803                                      dlm_get_lock_cookie_seq(c));
1804
1805                                 mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, "
1806                                      "node=%u, cookie=%u:%llu, queue=%d\n",
1807                                      ml->type, ml->convert_type, ml->node,
1808                                      dlm_get_lock_cookie_node(ml->cookie),
1809                                      dlm_get_lock_cookie_seq(ml->cookie),
1810                                      ml->list);
1811
1812                                 __dlm_print_one_lock_resource(res);
1813                                 bad = 1;
1814                                 break;
1815                         }
1816                 }
1817                 if (!bad) {
1818                         dlm_lock_get(newlock);
1819                         list_add_tail(&newlock->list, queue);
1820                 }
1821                 spin_unlock(&res->spinlock);
1822         }
1823         mlog(0, "done running all the locks\n");
1824
1825 leave:
1826         if (ret < 0) {
1827                 mlog_errno(ret);
1828                 if (newlock)
1829                         dlm_lock_put(newlock);
1830         }
1831
1832         mlog_exit(ret);
1833         return ret;
1834 }
1835
1836 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
1837                                        struct dlm_lock_resource *res)
1838 {
1839         int i;
1840         struct list_head *queue, *iter, *iter2;
1841         struct dlm_lock *lock;
1842
1843         res->state |= DLM_LOCK_RES_RECOVERING;
1844         if (!list_empty(&res->recovering)) {
1845                 mlog(0,
1846                      "Recovering res %s:%.*s, is already on recovery list!\n",
1847                      dlm->name, res->lockname.len, res->lockname.name);
1848                 list_del_init(&res->recovering);
1849         }
1850         /* We need to hold a reference while on the recovery list */
1851         dlm_lockres_get(res);
1852         list_add_tail(&res->recovering, &dlm->reco.resources);
1853
1854         /* find any pending locks and put them back on proper list */
1855         for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) {
1856                 queue = dlm_list_idx_to_ptr(res, i);
1857                 list_for_each_safe(iter, iter2, queue) {
1858                         lock = list_entry (iter, struct dlm_lock, list);
1859                         dlm_lock_get(lock);
1860                         if (lock->convert_pending) {
1861                                 /* move converting lock back to granted */
1862                                 BUG_ON(i != DLM_CONVERTING_LIST);
1863                                 mlog(0, "node died with convert pending "
1864                                      "on %.*s. move back to granted list.\n",
1865                                      res->lockname.len, res->lockname.name);
1866                                 dlm_revert_pending_convert(res, lock);
1867                                 lock->convert_pending = 0;
1868                         } else if (lock->lock_pending) {
1869                                 /* remove pending lock requests completely */
1870                                 BUG_ON(i != DLM_BLOCKED_LIST);
1871                                 mlog(0, "node died with lock pending "
1872                                      "on %.*s. remove from blocked list and skip.\n",
1873                                      res->lockname.len, res->lockname.name);
1874                                 /* lock will be floating until ref in
1875                                  * dlmlock_remote is freed after the network
1876                                  * call returns.  ok for it to not be on any
1877                                  * list since no ast can be called
1878                                  * (the master is dead). */
1879                                 dlm_revert_pending_lock(res, lock);
1880                                 lock->lock_pending = 0;
1881                         } else if (lock->unlock_pending) {
1882                                 /* if an unlock was in progress, treat as
1883                                  * if this had completed successfully
1884                                  * before sending this lock state to the
1885                                  * new master.  note that the dlm_unlock
1886                                  * call is still responsible for calling
1887                                  * the unlockast.  that will happen after
1888                                  * the network call times out.  for now,
1889                                  * just move lists to prepare the new
1890                                  * recovery master.  */
1891                                 BUG_ON(i != DLM_GRANTED_LIST);
1892                                 mlog(0, "node died with unlock pending "
1893                                      "on %.*s. remove from blocked list and skip.\n",
1894                                      res->lockname.len, res->lockname.name);
1895                                 dlm_commit_pending_unlock(res, lock);
1896                                 lock->unlock_pending = 0;
1897                         } else if (lock->cancel_pending) {
1898                                 /* if a cancel was in progress, treat as
1899                                  * if this had completed successfully
1900                                  * before sending this lock state to the
1901                                  * new master */
1902                                 BUG_ON(i != DLM_CONVERTING_LIST);
1903                                 mlog(0, "node died with cancel pending "
1904                                      "on %.*s. move back to granted list.\n",
1905                                      res->lockname.len, res->lockname.name);
1906                                 dlm_commit_pending_cancel(res, lock);
1907                                 lock->cancel_pending = 0;
1908                         }
1909                         dlm_lock_put(lock);
1910                 }
1911         }
1912 }
1913
1914
1915
1916 /* removes all recovered locks from the recovery list.
1917  * sets the res->owner to the new master.
1918  * unsets the RECOVERY flag and wakes waiters. */
1919 static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
1920                                               u8 dead_node, u8 new_master)
1921 {
1922         int i;
1923         struct list_head *iter, *iter2;
1924         struct hlist_node *hash_iter;
1925         struct hlist_head *bucket;
1926
1927         struct dlm_lock_resource *res;
1928
1929         mlog_entry_void();
1930
1931         assert_spin_locked(&dlm->spinlock);
1932
1933         list_for_each_safe(iter, iter2, &dlm->reco.resources) {
1934                 res = list_entry (iter, struct dlm_lock_resource, recovering);
1935                 if (res->owner == dead_node) {
1936                         list_del_init(&res->recovering);
1937                         spin_lock(&res->spinlock);
1938                         dlm_change_lockres_owner(dlm, res, new_master);
1939                         res->state &= ~DLM_LOCK_RES_RECOVERING;
1940                         if (!__dlm_lockres_unused(res))
1941                                 __dlm_dirty_lockres(dlm, res);
1942                         spin_unlock(&res->spinlock);
1943                         wake_up(&res->wq);
1944                         dlm_lockres_put(res);
1945                 }
1946         }
1947
1948         /* this will become unnecessary eventually, but
1949          * for now we need to run the whole hash, clear
1950          * the RECOVERING state and set the owner
1951          * if necessary */
1952         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
1953                 bucket = dlm_lockres_hash(dlm, i);
1954                 hlist_for_each_entry(res, hash_iter, bucket, hash_node) {
1955                         if (res->state & DLM_LOCK_RES_RECOVERING) {
1956                                 if (res->owner == dead_node) {
1957                                         mlog(0, "(this=%u) res %.*s owner=%u "
1958                                              "was not on recovering list, but "
1959                                              "clearing state anyway\n",
1960                                              dlm->node_num, res->lockname.len,
1961                                              res->lockname.name, new_master);
1962                                 } else if (res->owner == dlm->node_num) {
1963                                         mlog(0, "(this=%u) res %.*s owner=%u "
1964                                              "was not on recovering list, "
1965                                              "owner is THIS node, clearing\n",
1966                                              dlm->node_num, res->lockname.len,
1967                                              res->lockname.name, new_master);
1968                                 } else
1969                                         continue;
1970
1971                                 if (!list_empty(&res->recovering)) {
1972                                         mlog(0, "%s:%.*s: lockres was "
1973                                              "marked RECOVERING, owner=%u\n",
1974                                              dlm->name, res->lockname.len,
1975                                              res->lockname.name, res->owner);
1976                                         list_del_init(&res->recovering);
1977                                         dlm_lockres_put(res);
1978                                 }
1979                                 spin_lock(&res->spinlock);
1980                                 dlm_change_lockres_owner(dlm, res, new_master);
1981                                 res->state &= ~DLM_LOCK_RES_RECOVERING;
1982                                 if (!__dlm_lockres_unused(res))
1983                                         __dlm_dirty_lockres(dlm, res);
1984                                 spin_unlock(&res->spinlock);
1985                                 wake_up(&res->wq);
1986                         }
1987                 }
1988         }
1989 }
1990
1991 static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local)
1992 {
1993         if (local) {
1994                 if (lock->ml.type != LKM_EXMODE &&
1995                     lock->ml.type != LKM_PRMODE)
1996                         return 1;
1997         } else if (lock->ml.type == LKM_EXMODE)
1998                 return 1;
1999         return 0;
2000 }
2001
2002 static void dlm_revalidate_lvb(struct dlm_ctxt *dlm,
2003                                struct dlm_lock_resource *res, u8 dead_node)
2004 {
2005         struct list_head *iter, *queue;
2006         struct dlm_lock *lock;
2007         int blank_lvb = 0, local = 0;
2008         int i;
2009         u8 search_node;
2010
2011         assert_spin_locked(&dlm->spinlock);
2012         assert_spin_locked(&res->spinlock);
2013
2014         if (res->owner == dlm->node_num)
2015                 /* if this node owned the lockres, and if the dead node 
2016                  * had an EX when he died, blank out the lvb */
2017                 search_node = dead_node;
2018         else {
2019                 /* if this is a secondary lockres, and we had no EX or PR
2020                  * locks granted, we can no longer trust the lvb */
2021                 search_node = dlm->node_num;
2022                 local = 1;  /* check local state for valid lvb */
2023         }
2024
2025         for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) {
2026                 queue = dlm_list_idx_to_ptr(res, i);
2027                 list_for_each(iter, queue) {
2028                         lock = list_entry (iter, struct dlm_lock, list);
2029                         if (lock->ml.node == search_node) {
2030                                 if (dlm_lvb_needs_invalidation(lock, local)) {
2031                                         /* zero the lksb lvb and lockres lvb */
2032                                         blank_lvb = 1;
2033                                         memset(lock->lksb->lvb, 0, DLM_LVB_LEN);
2034                                 }
2035                         }
2036                 }
2037         }
2038
2039         if (blank_lvb) {
2040                 mlog(0, "clearing %.*s lvb, dead node %u had EX\n",
2041                      res->lockname.len, res->lockname.name, dead_node);
2042                 memset(res->lvb, 0, DLM_LVB_LEN);
2043         }
2044 }
2045
2046 static void dlm_free_dead_locks(struct dlm_ctxt *dlm,
2047                                 struct dlm_lock_resource *res, u8 dead_node)
2048 {
2049         struct list_head *iter, *tmpiter;
2050         struct dlm_lock *lock;
2051
2052         /* this node is the lockres master:
2053          * 1) remove any stale locks for the dead node
2054          * 2) if the dead node had an EX when he died, blank out the lvb 
2055          */
2056         assert_spin_locked(&dlm->spinlock);
2057         assert_spin_locked(&res->spinlock);
2058
2059         /* TODO: check pending_asts, pending_basts here */
2060         list_for_each_safe(iter, tmpiter, &res->granted) {
2061                 lock = list_entry (iter, struct dlm_lock, list);
2062                 if (lock->ml.node == dead_node) {
2063                         list_del_init(&lock->list);
2064                         dlm_lock_put(lock);
2065                 }
2066         }
2067         list_for_each_safe(iter, tmpiter, &res->converting) {
2068                 lock = list_entry (iter, struct dlm_lock, list);
2069                 if (lock->ml.node == dead_node) {
2070                         list_del_init(&lock->list);
2071                         dlm_lock_put(lock);
2072                 }
2073         }
2074         list_for_each_safe(iter, tmpiter, &res->blocked) {
2075                 lock = list_entry (iter, struct dlm_lock, list);
2076                 if (lock->ml.node == dead_node) {
2077                         list_del_init(&lock->list);
2078                         dlm_lock_put(lock);
2079                 }
2080         }
2081
2082         /* do not kick thread yet */
2083         __dlm_dirty_lockres(dlm, res);
2084 }
2085
2086 /* if this node is the recovery master, and there are no
2087  * locks for a given lockres owned by this node that are in
2088  * either PR or EX mode, zero out the lvb before requesting.
2089  *
2090  */
2091
2092
2093 static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
2094 {
2095         struct hlist_node *iter;
2096         struct dlm_lock_resource *res;
2097         int i;
2098         struct hlist_head *bucket;
2099         struct dlm_lock *lock;
2100
2101
2102         /* purge any stale mles */
2103         dlm_clean_master_list(dlm, dead_node);
2104
2105         /*
2106          * now clean up all lock resources.  there are two rules:
2107          *
2108          * 1) if the dead node was the master, move the lockres
2109          *    to the recovering list.  set the RECOVERING flag.
2110          *    this lockres needs to be cleaned up before it can
2111          *    be used further.
2112          *
2113          * 2) if this node was the master, remove all locks from
2114          *    each of the lockres queues that were owned by the
2115          *    dead node.  once recovery finishes, the dlm thread
2116          *    can be kicked again to see if any ASTs or BASTs
2117          *    need to be fired as a result.
2118          */
2119         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
2120                 bucket = dlm_lockres_hash(dlm, i);
2121                 hlist_for_each_entry(res, iter, bucket, hash_node) {
2122                         /* always prune any $RECOVERY entries for dead nodes,
2123                          * otherwise hangs can occur during later recovery */
2124                         if (dlm_is_recovery_lock(res->lockname.name,
2125                                                  res->lockname.len)) {
2126                                 spin_lock(&res->spinlock);
2127                                 list_for_each_entry(lock, &res->granted, list) {
2128                                         if (lock->ml.node == dead_node) {
2129                                                 mlog(0, "AHA! there was "
2130                                                      "a $RECOVERY lock for dead "
2131                                                      "node %u (%s)!\n",
2132                                                      dead_node, dlm->name);
2133                                                 list_del_init(&lock->list);
2134                                                 dlm_lock_put(lock);
2135                                                 break;
2136                                         }
2137                                 }
2138                                 spin_unlock(&res->spinlock);
2139                                 continue;
2140                         }                       
2141                         spin_lock(&res->spinlock);
2142                         /* zero the lvb if necessary */
2143                         dlm_revalidate_lvb(dlm, res, dead_node);
2144                         if (res->owner == dead_node)
2145                                 dlm_move_lockres_to_recovery_list(dlm, res);
2146                         else if (res->owner == dlm->node_num) {
2147                                 dlm_free_dead_locks(dlm, res, dead_node);
2148                                 __dlm_lockres_calc_usage(dlm, res);
2149                         }
2150                         spin_unlock(&res->spinlock);
2151                 }
2152         }
2153
2154 }
2155
2156 static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx)
2157 {
2158         assert_spin_locked(&dlm->spinlock);
2159
2160         if (dlm->reco.new_master == idx) {
2161                 mlog(0, "%s: recovery master %d just died\n",
2162                      dlm->name, idx);
2163                 if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2164                         /* finalize1 was reached, so it is safe to clear
2165                          * the new_master and dead_node.  that recovery
2166                          * is complete. */
2167                         mlog(0, "%s: dead master %d had reached "
2168                              "finalize1 state, clearing\n", dlm->name, idx);
2169                         dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2170                         __dlm_reset_recovery(dlm);
2171                 }
2172         }
2173
2174         /* check to see if the node is already considered dead */
2175         if (!test_bit(idx, dlm->live_nodes_map)) {
2176                 mlog(0, "for domain %s, node %d is already dead. "
2177                      "another node likely did recovery already.\n",
2178                      dlm->name, idx);
2179                 return;
2180         }
2181
2182         /* check to see if we do not care about this node */
2183         if (!test_bit(idx, dlm->domain_map)) {
2184                 /* This also catches the case that we get a node down
2185                  * but haven't joined the domain yet. */
2186                 mlog(0, "node %u already removed from domain!\n", idx);
2187                 return;
2188         }
2189
2190         clear_bit(idx, dlm->live_nodes_map);
2191
2192         /* Clean up join state on node death. */
2193         if (dlm->joining_node == idx) {
2194                 mlog(0, "Clearing join state for node %u\n", idx);
2195                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
2196         }
2197
2198         /* make sure local cleanup occurs before the heartbeat events */
2199         if (!test_bit(idx, dlm->recovery_map))
2200                 dlm_do_local_recovery_cleanup(dlm, idx);
2201
2202         /* notify anything attached to the heartbeat events */
2203         dlm_hb_event_notify_attached(dlm, idx, 0);
2204
2205         mlog(0, "node %u being removed from domain map!\n", idx);
2206         clear_bit(idx, dlm->domain_map);
2207         /* wake up migration waiters if a node goes down.
2208          * perhaps later we can genericize this for other waiters. */
2209         wake_up(&dlm->migration_wq);
2210
2211         if (test_bit(idx, dlm->recovery_map))
2212                 mlog(0, "domain %s, node %u already added "
2213                      "to recovery map!\n", dlm->name, idx);
2214         else
2215                 set_bit(idx, dlm->recovery_map);
2216 }
2217
2218 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data)
2219 {
2220         struct dlm_ctxt *dlm = data;
2221
2222         if (!dlm_grab(dlm))
2223                 return;
2224
2225         spin_lock(&dlm->spinlock);
2226         __dlm_hb_node_down(dlm, idx);
2227         spin_unlock(&dlm->spinlock);
2228
2229         dlm_put(dlm);
2230 }
2231
2232 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data)
2233 {
2234         struct dlm_ctxt *dlm = data;
2235
2236         if (!dlm_grab(dlm))
2237                 return;
2238
2239         spin_lock(&dlm->spinlock);
2240         set_bit(idx, dlm->live_nodes_map);
2241         /* do NOT notify mle attached to the heartbeat events.
2242          * new nodes are not interesting in mastery until joined. */
2243         spin_unlock(&dlm->spinlock);
2244
2245         dlm_put(dlm);
2246 }
2247
2248 static void dlm_reco_ast(void *astdata)
2249 {
2250         struct dlm_ctxt *dlm = astdata;
2251         mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n",
2252              dlm->node_num, dlm->name);
2253 }
2254 static void dlm_reco_bast(void *astdata, int blocked_type)
2255 {
2256         struct dlm_ctxt *dlm = astdata;
2257         mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n",
2258              dlm->node_num, dlm->name);
2259 }
2260 static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st)
2261 {
2262         mlog(0, "unlockast for recovery lock fired!\n");
2263 }
2264
2265 /*
2266  * dlm_pick_recovery_master will continually attempt to use
2267  * dlmlock() on the special "$RECOVERY" lockres with the
2268  * LKM_NOQUEUE flag to get an EX.  every thread that enters
2269  * this function on each node racing to become the recovery
2270  * master will not stop attempting this until either:
2271  * a) this node gets the EX (and becomes the recovery master),
2272  * or b) dlm->reco.new_master gets set to some nodenum 
2273  * != O2NM_INVALID_NODE_NUM (another node will do the reco).
2274  * so each time a recovery master is needed, the entire cluster
2275  * will sync at this point.  if the new master dies, that will
2276  * be detected in dlm_do_recovery */
2277 static int dlm_pick_recovery_master(struct dlm_ctxt *dlm)
2278 {
2279         enum dlm_status ret;
2280         struct dlm_lockstatus lksb;
2281         int status = -EINVAL;
2282
2283         mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
2284              dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num);
2285 again:  
2286         memset(&lksb, 0, sizeof(lksb));
2287
2288         ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY,
2289                       DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN,
2290                       dlm_reco_ast, dlm, dlm_reco_bast);
2291
2292         mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n",
2293              dlm->name, ret, lksb.status);
2294
2295         if (ret == DLM_NORMAL) {
2296                 mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
2297                      dlm->name, dlm->node_num);
2298                 
2299                 /* got the EX lock.  check to see if another node 
2300                  * just became the reco master */
2301                 if (dlm_reco_master_ready(dlm)) {
2302                         mlog(0, "%s: got reco EX lock, but %u will "
2303                              "do the recovery\n", dlm->name,
2304                              dlm->reco.new_master);
2305                         status = -EEXIST;
2306                 } else {
2307                         status = 0;
2308
2309                         /* see if recovery was already finished elsewhere */
2310                         spin_lock(&dlm->spinlock);
2311                         if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
2312                                 status = -EINVAL;       
2313                                 mlog(0, "%s: got reco EX lock, but "
2314                                      "node got recovered already\n", dlm->name);
2315                                 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2316                                         mlog(ML_ERROR, "%s: new master is %u "
2317                                              "but no dead node!\n", 
2318                                              dlm->name, dlm->reco.new_master);
2319                                         BUG();
2320                                 }
2321                         }
2322                         spin_unlock(&dlm->spinlock);
2323                 }
2324
2325                 /* if this node has actually become the recovery master,
2326                  * set the master and send the messages to begin recovery */
2327                 if (!status) {
2328                         mlog(0, "%s: dead=%u, this=%u, sending "
2329                              "begin_reco now\n", dlm->name, 
2330                              dlm->reco.dead_node, dlm->node_num);
2331                         status = dlm_send_begin_reco_message(dlm,
2332                                       dlm->reco.dead_node);
2333                         /* this always succeeds */
2334                         BUG_ON(status);
2335
2336                         /* set the new_master to this node */
2337                         spin_lock(&dlm->spinlock);
2338                         dlm_set_reco_master(dlm, dlm->node_num);
2339                         spin_unlock(&dlm->spinlock);
2340                 }
2341
2342                 /* recovery lock is a special case.  ast will not get fired,
2343                  * so just go ahead and unlock it. */
2344                 ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm);
2345                 if (ret == DLM_DENIED) {
2346                         mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n");
2347                         ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm);
2348                 }
2349                 if (ret != DLM_NORMAL) {
2350                         /* this would really suck. this could only happen
2351                          * if there was a network error during the unlock
2352                          * because of node death.  this means the unlock
2353                          * is actually "done" and the lock structure is
2354                          * even freed.  we can continue, but only
2355                          * because this specific lock name is special. */
2356                         mlog(ML_ERROR, "dlmunlock returned %d\n", ret);
2357                 }
2358         } else if (ret == DLM_NOTQUEUED) {
2359                 mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
2360                      dlm->name, dlm->node_num);
2361                 /* another node is master. wait on
2362                  * reco.new_master != O2NM_INVALID_NODE_NUM 
2363                  * for at most one second */
2364                 wait_event_timeout(dlm->dlm_reco_thread_wq,
2365                                          dlm_reco_master_ready(dlm),
2366                                          msecs_to_jiffies(1000));
2367                 if (!dlm_reco_master_ready(dlm)) {
2368                         mlog(0, "%s: reco master taking awhile\n",
2369                              dlm->name);
2370                         goto again;
2371                 }
2372                 /* another node has informed this one that it is reco master */
2373                 mlog(0, "%s: reco master %u is ready to recover %u\n",
2374                      dlm->name, dlm->reco.new_master, dlm->reco.dead_node);
2375                 status = -EEXIST;
2376         } else if (ret == DLM_RECOVERING) {
2377                 mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n",
2378                      dlm->name, dlm->node_num);
2379                 goto again;
2380         } else {
2381                 struct dlm_lock_resource *res;
2382
2383                 /* dlmlock returned something other than NOTQUEUED or NORMAL */
2384                 mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), "
2385                      "lksb.status=%s\n", dlm->name, dlm_errname(ret),
2386                      dlm_errname(lksb.status));
2387                 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2388                                          DLM_RECOVERY_LOCK_NAME_LEN);
2389                 if (res) {
2390                         dlm_print_one_lock_resource(res);
2391                         dlm_lockres_put(res);
2392                 } else {
2393                         mlog(ML_ERROR, "recovery lock not found\n");
2394                 }
2395                 BUG();
2396         }
2397
2398         return status;
2399 }
2400
2401 static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node)
2402 {
2403         struct dlm_begin_reco br;
2404         int ret = 0;
2405         struct dlm_node_iter iter;
2406         int nodenum;
2407         int status;
2408
2409         mlog_entry("%u\n", dead_node);
2410
2411         mlog(0, "%s: dead node is %u\n", dlm->name, dead_node);
2412
2413         spin_lock(&dlm->spinlock);
2414         dlm_node_iter_init(dlm->domain_map, &iter);
2415         spin_unlock(&dlm->spinlock);
2416
2417         clear_bit(dead_node, iter.node_map);
2418
2419         memset(&br, 0, sizeof(br));
2420         br.node_idx = dlm->node_num;
2421         br.dead_node = dead_node;
2422
2423         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2424                 ret = 0;
2425                 if (nodenum == dead_node) {
2426                         mlog(0, "not sending begin reco to dead node "
2427                                   "%u\n", dead_node);
2428                         continue;
2429                 }
2430                 if (nodenum == dlm->node_num) {
2431                         mlog(0, "not sending begin reco to self\n");
2432                         continue;
2433                 }
2434 retry:
2435                 ret = -EINVAL;
2436                 mlog(0, "attempting to send begin reco msg to %d\n",
2437                           nodenum);
2438                 ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key,
2439                                          &br, sizeof(br), nodenum, &status);
2440                 /* negative status is handled ok by caller here */
2441                 if (ret >= 0)
2442                         ret = status;
2443                 if (dlm_is_host_down(ret)) {
2444                         /* node is down.  not involved in recovery
2445                          * so just keep going */
2446                         mlog(0, "%s: node %u was down when sending "
2447                              "begin reco msg (%d)\n", dlm->name, nodenum, ret);
2448                         ret = 0;
2449                 }
2450                 if (ret < 0) {
2451                         struct dlm_lock_resource *res;
2452                         /* this is now a serious problem, possibly ENOMEM 
2453                          * in the network stack.  must retry */
2454                         mlog_errno(ret);
2455                         mlog(ML_ERROR, "begin reco of dlm %s to node %u "
2456                             " returned %d\n", dlm->name, nodenum, ret);
2457                         res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2458                                                  DLM_RECOVERY_LOCK_NAME_LEN);
2459                         if (res) {
2460                                 dlm_print_one_lock_resource(res);
2461                                 dlm_lockres_put(res);
2462                         } else {
2463                                 mlog(ML_ERROR, "recovery lock not found\n");
2464                         }
2465                         /* sleep for a bit in hopes that we can avoid 
2466                          * another ENOMEM */
2467                         msleep(100);
2468                         goto retry;
2469                 } else if (ret == EAGAIN) {
2470                         mlog(0, "%s: trying to start recovery of node "
2471                              "%u, but node %u is waiting for last recovery "
2472                              "to complete, backoff for a bit\n", dlm->name,
2473                              dead_node, nodenum);
2474                         /* TODO Look into replacing msleep with cond_resched() */
2475                         msleep(100);
2476                         goto retry;
2477                 }
2478         }
2479
2480         return ret;
2481 }
2482
2483 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data)
2484 {
2485         struct dlm_ctxt *dlm = data;
2486         struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf;
2487
2488         /* ok to return 0, domain has gone away */
2489         if (!dlm_grab(dlm))
2490                 return 0;
2491
2492         spin_lock(&dlm->spinlock);
2493         if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2494                 mlog(0, "%s: node %u wants to recover node %u (%u:%u) "
2495                      "but this node is in finalize state, waiting on finalize2\n",
2496                      dlm->name, br->node_idx, br->dead_node,
2497                      dlm->reco.dead_node, dlm->reco.new_master);
2498                 spin_unlock(&dlm->spinlock);
2499                 return EAGAIN;
2500         }
2501         spin_unlock(&dlm->spinlock);
2502
2503         mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n",
2504              dlm->name, br->node_idx, br->dead_node,
2505              dlm->reco.dead_node, dlm->reco.new_master);
2506
2507         dlm_fire_domain_eviction_callbacks(dlm, br->dead_node);
2508
2509         spin_lock(&dlm->spinlock);
2510         if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2511                 if (test_bit(dlm->reco.new_master, dlm->recovery_map)) {
2512                         mlog(0, "%s: new_master %u died, changing "
2513                              "to %u\n", dlm->name, dlm->reco.new_master,
2514                              br->node_idx);
2515                 } else {
2516                         mlog(0, "%s: new_master %u NOT DEAD, changing "
2517                              "to %u\n", dlm->name, dlm->reco.new_master,
2518                              br->node_idx);
2519                         /* may not have seen the new master as dead yet */
2520                 }
2521         }
2522         if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) {
2523                 mlog(ML_NOTICE, "%s: dead_node previously set to %u, "
2524                      "node %u changing it to %u\n", dlm->name, 
2525                      dlm->reco.dead_node, br->node_idx, br->dead_node);
2526         }
2527         dlm_set_reco_master(dlm, br->node_idx);
2528         dlm_set_reco_dead_node(dlm, br->dead_node);
2529         if (!test_bit(br->dead_node, dlm->recovery_map)) {
2530                 mlog(0, "recovery master %u sees %u as dead, but this "
2531                      "node has not yet.  marking %u as dead\n",
2532                      br->node_idx, br->dead_node, br->dead_node);
2533                 if (!test_bit(br->dead_node, dlm->domain_map) ||
2534                     !test_bit(br->dead_node, dlm->live_nodes_map))
2535                         mlog(0, "%u not in domain/live_nodes map "
2536                              "so setting it in reco map manually\n",
2537                              br->dead_node);
2538                 /* force the recovery cleanup in __dlm_hb_node_down
2539                  * both of these will be cleared in a moment */
2540                 set_bit(br->dead_node, dlm->domain_map);
2541                 set_bit(br->dead_node, dlm->live_nodes_map);
2542                 __dlm_hb_node_down(dlm, br->dead_node);
2543         }
2544         spin_unlock(&dlm->spinlock);
2545
2546         dlm_kick_recovery_thread(dlm);
2547
2548         mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n",
2549              dlm->name, br->node_idx, br->dead_node,
2550              dlm->reco.dead_node, dlm->reco.new_master);
2551
2552         dlm_put(dlm);
2553         return 0;
2554 }
2555
2556 #define DLM_FINALIZE_STAGE2  0x01
2557 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
2558 {
2559         int ret = 0;
2560         struct dlm_finalize_reco fr;
2561         struct dlm_node_iter iter;
2562         int nodenum;
2563         int status;
2564         int stage = 1;
2565
2566         mlog(0, "finishing recovery for node %s:%u, "
2567              "stage %d\n", dlm->name, dlm->reco.dead_node, stage);
2568
2569         spin_lock(&dlm->spinlock);
2570         dlm_node_iter_init(dlm->domain_map, &iter);
2571         spin_unlock(&dlm->spinlock);
2572
2573 stage2:
2574         memset(&fr, 0, sizeof(fr));
2575         fr.node_idx = dlm->node_num;
2576         fr.dead_node = dlm->reco.dead_node;
2577         if (stage == 2)
2578                 fr.flags |= DLM_FINALIZE_STAGE2;
2579
2580         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2581                 if (nodenum == dlm->node_num)
2582                         continue;
2583                 ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key,
2584                                          &fr, sizeof(fr), nodenum, &status);
2585                 if (ret >= 0)
2586                         ret = status;
2587                 if (ret < 0) {
2588                         mlog_errno(ret);
2589                         if (dlm_is_host_down(ret)) {
2590                                 /* this has no effect on this recovery 
2591                                  * session, so set the status to zero to 
2592                                  * finish out the last recovery */
2593                                 mlog(ML_ERROR, "node %u went down after this "
2594                                      "node finished recovery.\n", nodenum);
2595                                 ret = 0;
2596                                 continue;
2597                         }
2598                         break;
2599                 }
2600         }
2601         if (stage == 1) {
2602                 /* reset the node_iter back to the top and send finalize2 */
2603                 iter.curnode = -1;
2604                 stage = 2;
2605                 goto stage2;
2606         }
2607
2608         return ret;
2609 }
2610
2611 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data)
2612 {
2613         struct dlm_ctxt *dlm = data;
2614         struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf;
2615         int stage = 1;
2616
2617         /* ok to return 0, domain has gone away */
2618         if (!dlm_grab(dlm))
2619                 return 0;
2620
2621         if (fr->flags & DLM_FINALIZE_STAGE2)
2622                 stage = 2;
2623
2624         mlog(0, "%s: node %u finalizing recovery stage%d of "
2625              "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage,
2626              fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master);
2627  
2628         spin_lock(&dlm->spinlock);
2629
2630         if (dlm->reco.new_master != fr->node_idx) {
2631                 mlog(ML_ERROR, "node %u sent recovery finalize msg, but node "
2632                      "%u is supposed to be the new master, dead=%u\n",
2633                      fr->node_idx, dlm->reco.new_master, fr->dead_node);
2634                 BUG();
2635         }
2636         if (dlm->reco.dead_node != fr->dead_node) {
2637                 mlog(ML_ERROR, "node %u sent recovery finalize msg for dead "
2638                      "node %u, but node %u is supposed to be dead\n",
2639                      fr->node_idx, fr->dead_node, dlm->reco.dead_node);
2640                 BUG();
2641         }
2642
2643         switch (stage) {
2644                 case 1:
2645                         dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
2646                         if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
2647                                 mlog(ML_ERROR, "%s: received finalize1 from "
2648                                      "new master %u for dead node %u, but "
2649                                      "this node has already received it!\n",
2650                                      dlm->name, fr->node_idx, fr->dead_node);
2651                                 dlm_print_reco_node_status(dlm);
2652                                 BUG();
2653                         }
2654                         dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
2655                         spin_unlock(&dlm->spinlock);
2656                         break;
2657                 case 2:
2658                         if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) {
2659                                 mlog(ML_ERROR, "%s: received finalize2 from "
2660                                      "new master %u for dead node %u, but "
2661                                      "this node did not have finalize1!\n",
2662                                      dlm->name, fr->node_idx, fr->dead_node);
2663                                 dlm_print_reco_node_status(dlm);
2664                                 BUG();
2665                         }
2666                         dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
2667                         spin_unlock(&dlm->spinlock);
2668                         dlm_reset_recovery(dlm);
2669                         dlm_kick_recovery_thread(dlm);
2670                         break;
2671                 default:
2672                         BUG();
2673         }
2674
2675         mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n",
2676              dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master);
2677
2678         dlm_put(dlm);
2679         return 0;
2680 }